English | [简体中文](README_ch.md) - [Document Visual Question Answering (Doc-VQA)](#Document-Visual-Question-Answering) - [1. Introduction](#1-Introduction) - [2. Performance](#2-performance) - [3. Effect demo](#3-Effect-demo) - [3.1 SER](#31-ser) - [3.2 RE](#32-re) - [4. Install](#4-Install) - [4.1 Installation dependencies](#41-Install-dependencies) - [4.2 Install PaddleOCR](#42-Install-PaddleOCR) - [5. Usage](#5-Usage) - [5.1 Data and Model Preparation](#51-Data-and-Model-Preparation) - [5.2 SER](#52-ser) - [5.3 RE](#53-re) - [6. Reference](#6-Reference-Links) # Document Visual Question Answering ## 1 Introduction VQA refers to visual question answering, which mainly asks and answers image content. DOC-VQA is one of the VQA tasks. DOC-VQA mainly asks questions about the text content of text images. The DOC-VQA algorithm in PP-Structure is developed based on the PaddleNLP natural language processing algorithm library. The main features are as follows: - Integrate [LayoutXLM](https://arxiv.org/pdf/2104.08836.pdf) model and PP-OCR prediction engine. - Supports Semantic Entity Recognition (SER) and Relation Extraction (RE) tasks based on multimodal methods. Based on the SER task, the text recognition and classification in the image can be completed; based on the RE task, the relationship extraction of the text content in the image can be completed, such as judging the problem pair (pair). - Supports custom training for SER tasks and RE tasks. - Supports end-to-end system prediction and evaluation of OCR+SER. - Supports end-to-end system prediction of OCR+SER+RE. This project is an open source implementation of [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/pdf/2104.08836.pdf) on Paddle 2.2, Included fine-tuning code on [XFUND dataset](https://github.com/doc-analysis/XFUND). ## 2. Performance We evaluate the algorithm on the Chinese dataset of [XFUND](https://github.com/doc-analysis/XFUND), and the performance is as follows | Model | Task | hmean | Model download address | |:---:|:---:|:---:| :---:| | LayoutXLM | SER | 0.9038 | [link](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) | | LayoutXLM | RE | 0.7483 | [link](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) | | LayoutLMv2 | SER | 0.8544 | [link](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar) | LayoutLMv2 | RE | 0.6777 | [link](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar) | | LayoutLM | SER | 0.7731 | [link](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar) | ## 3. Effect demo **Note:** The test images are from the XFUND dataset. ### 3.1 SER ![](../docs/vqa/result_ser/zh_val_0_ser.jpg) | ![](../docs/vqa/result_ser/zh_val_42_ser.jpg) ---|--- Boxes with different colors in the figure represent different categories. For the XFUND dataset, there are 3 categories: `QUESTION`, `ANSWER`, `HEADER` * Dark purple: HEADER * Light purple: QUESTION * Army Green: ANSWER The corresponding categories and OCR recognition results are also marked on the upper left of the OCR detection frame. ### 3.2 RE ![](../docs/vqa/result_re/zh_val_21_re.jpg) | ![](../docs/vqa/result_re/zh_val_40_re.jpg) ---|--- The red box in the figure represents the question, the blue box represents the answer, and the question and the answer are connected by a green line. The corresponding categories and OCR recognition results are also marked on the upper left of the OCR detection frame. ## 4. Install ### 4.1 Install dependencies - **(1) Install PaddlePaddle** ```bash python3 -m pip install --upgrade pip # GPU installation python3 -m pip install "paddlepaddle-gpu>=2.2" -i https://mirror.baidu.com/pypi/simple # CPU installation python3 -m pip install "paddlepaddle>=2.2" -i https://mirror.baidu.com/pypi/simple ```` For more requirements, please refer to the instructions in [Installation Documentation](https://www.paddlepaddle.org.cn/install/quick). ### 4.2 Install PaddleOCR - **(1) pip install PaddleOCR whl package quickly (prediction only)** ```bash python3 -m pip install paddleocr ```` - **(2) Download VQA source code (prediction + training)** ```bash [Recommended] git clone https://github.com/PaddlePaddle/PaddleOCR # If the pull cannot be successful due to network problems, you can also choose to use the hosting on the code cloud: git clone https://gitee.com/paddlepaddle/PaddleOCR # Note: Code cloud hosting code may not be able to synchronize the update of this github project in real time, there is a delay of 3 to 5 days, please use the recommended method first. ```` - **(3) Install VQA's `requirements`** ```bash python3 -m pip install -r ppstructure/vqa/requirements.txt ```` ## 5. Usage ### 5.1 Data and Model Preparation If you want to experience the prediction process directly, you can download the pre-training model provided by us, skip the training process, and just predict directly. * Download the processed dataset The download address of the processed XFUND Chinese dataset: [https://paddleocr.bj.bcebos.com/dataset/XFUND.tar](https://paddleocr.bj.bcebos.com/dataset/XFUND.tar). Download and unzip the dataset, and place the dataset in the current directory after unzipping. ```shell wget https://paddleocr.bj.bcebos.com/dataset/XFUND.tar ```` * Convert the dataset If you need to train other XFUND datasets, you can use the following commands to convert the datasets ```bash python3 ppstructure/vqa/tools/trans_xfun_data.py --ori_gt_path=path/to/json_path --output_path=path/to/save_path ```` * Download the pretrained models ```bash mkdir pretrain && cd pretrain #download the SER model wget https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar && tar -xvf ser_LayoutXLM_xfun_zh.tar #download the RE model wget https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar && tar -xvf re_LayoutXLM_xfun_zh.tar cd ../ ```` ### 5.2 SER Before starting training, you need to modify the following four fields 1. `Train.dataset.data_dir`: point to the directory where the training set images are stored 2. `Train.dataset.label_file_list`: point to the training set label file 3. `Eval.dataset.data_dir`: refers to the directory where the validation set images are stored 4. `Eval.dataset.label_file_list`: point to the validation set label file * start training ```shell CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/ser/layoutxlm.yml ```` Finally, `precision`, `recall`, `hmean` and other indicators will be printed. In the `./output/ser_layoutxlm/` folder will save the training log, the optimal model and the model for the latest epoch. * resume training To resume training, assign the folder path of the previously trained model to the `Architecture.Backbone.checkpoints` field. ```shell CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir ```` * evaluate Evaluation requires assigning the folder path of the model to be evaluated to the `Architecture.Backbone.checkpoints` field. ```shell CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir ```` Finally, `precision`, `recall`, `hmean` and other indicators will be printed * Use `OCR engine + SER` tandem prediction Use the following command to complete the series prediction of `OCR engine + SER`, taking the pretrained SER model as an example: ```shell CUDA_VISIBLE_DEVICES=0 python3 tools/infer_vqa_token_ser.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/Global.infer_img=doc/vqa/input/zh_val_42.jpg ```` Finally, the prediction result visualization image and the prediction result text file will be saved in the directory configured by the `config.Global.save_res_path` field. The prediction result text file is named `infer_results.txt`. * End-to-end evaluation of `OCR engine + SER` prediction system First use the `tools/infer_vqa_token_ser.py` script to complete the prediction of the dataset, then use the following command to evaluate. ```shell export CUDA_VISIBLE_DEVICES=0 python3 tools/eval_with_label_end2end.py --gt_json_path XFUND/zh_val/xfun_normalize_val.json --pred_json_path output_res/infer_results.txt ```` ### 5.3 RE * start training Before starting training, you need to modify the following four fields 1. `Train.dataset.data_dir`: point to the directory where the training set images are stored 2. `Train.dataset.label_file_list`: point to the training set label file 3. `Eval.dataset.data_dir`: refers to the directory where the validation set images are stored 4. `Eval.dataset.label_file_list`: point to the validation set label file ```shell CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/re/layoutxlm.yml ```` Finally, `precision`, `recall`, `hmean` and other indicators will be printed. In the `./output/re_layoutxlm/` folder will save the training log, the optimal model and the model for the latest epoch. * resume training To resume training, assign the folder path of the previously trained model to the `Architecture.Backbone.checkpoints` field. ```shell CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir ```` * evaluate Evaluation requires assigning the folder path of the model to be evaluated to the `Architecture.Backbone.checkpoints` field. ```shell CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir ```` Finally, `precision`, `recall`, `hmean` and other indicators will be printed * Use `OCR engine + SER + RE` tandem prediction Use the following command to complete the series prediction of `OCR engine + SER + RE`, taking the pretrained SER and RE models as an example: ```shell export CUDA_VISIBLE_DEVICES=0 python3 tools/infer_vqa_token_ser_re.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=pretrain/re_LayoutXLM_xfun_zh/Global.infer_img=doc/vqa/input/zh_val_21.jpg -c_ser configs/vqa/ser/layoutxlm. yml -o_ser Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/ ```` Finally, the prediction result visualization image and the prediction result text file will be saved in the directory configured by the `config.Global.save_res_path` field. The prediction result text file is named `infer_results.txt`. ## 6. Reference Links - LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding, https://arxiv.org/pdf/2104.08836.pdf - microsoft/unilm/layoutxlm, https://github.com/microsoft/unilm/tree/master/layoutxlm - XFUND dataset, https://github.com/doc-analysis/XFUND ## License The content of this project itself is licensed under the [Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/)